Statistical methods for modeling and nowcasting the impacts of influenza epidemics

Abstract

Influenza is an acute respiratory infection responsible for epidemics with high impact on human health. Several statistical methods have been applied to data collected from influenza surveillance systems (ISS) to assess the epidemic burden and early detect it. Given the ISS reporting delays, models have recently been developed to correct them by predicting the present situation (nowcasting) using the incomplete information collected. Thus, three objectives were defined. Review and classify the methods that use interrupted mortality time series to estimate influenza excess deaths. They were classified according to the model used to fit the time series and obtain a baseline; the influenza epidemic period estimator and the procedure used to fit the model (iterative or non iterative). This generalization led to the development of user friendly R-package, flubase, implementing all these models. Estimate influenza excess deaths in Portugal between 1980 and 2004. The seasonal excess deaths average by all causes was 2,475, of those 90% occurred in the elderly. These results suggest a similar influenza epidemics profile between Portugal and other countries in the Northern Hemisphere, and represent the first reference to contextualize future epidemics severity and design public health measures. Develop a model to nowcast the influenza epidemic evolution in a weekly basis. A non homogenous hidden Markov model (HMM) was developed to nowcast the current week influenza-like illness (ILI) incidence rate and the probability that the influenza activity is epidemic using as covariates an early estimate of ILI rate and the number of ILI cases tested positive in the previous week. Bayesian inference was used to estimate the model parameters and nowcasted quantities. The results obtained by application to the Portuguese ISS data, demonstrated the additional value of using a non homogenous HMM instead of an homogenous since it improves the ISS timeliness in 2 weeks.